标题： 加速多模型贝叶斯推断、模型选择和引力波天文学的系统研究 显示英文标题

标题： Accelerating Multi-Model Bayesian Inference, Model Selection and Systematic Studies for Gravitational Wave Astronomy

摘要： 引力波模型通过贝叶斯推断用于从观测到的引力波信号中推断合并双星系统中黑洞的属性。 尽管我们拥有大量足够准确的信号模型，可用于推断黑洞的属性，但对于某些信号，模型中的微小差异会导致推断属性的系统性差异。 为了对黑洞属性提供一个单一的估计，最好是将模型不确定性进行边缘化处理。 贝叶斯模型平均是一种常用的边缘化多个模型的技术，然而，它计算成本较高。 一种优雅的解决方案是通过联合贝叶斯分析同时推断模型和模型属性。 在本工作中，我们证明联合贝叶斯分析不仅可以加速推断，还可以考虑推断出的黑洞属性中的模型依赖性系统差异。 我们通过分析100个随机选择的模拟信号以及真实的引力波信号GW200129_065458来验证这一技术。 我们发现，我们不仅能够推断出与使用贝叶斯模型平均所得属性统计上相同的属性，而且平均而言，我们可以以$2.5\times$的速度对一组三个模型进行采样。 换句话说，对三个模型进行边缘化的联合贝叶斯分析，平均仅比单个模型分析多花费$20\%$的时间。 然后，我们展示了该技术可以用来准确且高效地量化一个模型相对于另一个模型的支持程度，从而有助于贝叶斯模型选择。 显示英文摘要

摘要： Gravitational wave models are used to infer the properties of black holes in merging binaries from the observed gravitational wave signals through Bayesian inference. Although we have access to a large collection of signal models that are sufficiently accurate to infer the properties of black holes, for some signals, small discrepancies in the models lead to systematic differences in the inferred properties. In order to provide a single estimate for the properties of the black holes, it is preferable to marginalize over the model uncertainty. Bayesian model averaging is a commonly used technique to marginalize over multiple models, however, it is computationally expensive. An elegant solution is to simultaneously infer the model and model properties in a joint Bayesian analysis. In this work we demonstrate that a joint Bayesian analysis can not only accelerate but also account for model-dependent systematic differences in the inferred black hole properties. We verify this technique by analysing 100 randomly chosen simulated signals and also the real gravitational wave signal GW200129_065458. We find that not only do we infer statistically identical properties as those obtained using Bayesian model averaging, but we can sample over a set of three models on average $2.5\times$ faster. In other words, a joint Bayesian analysis that marginalizes over three models takes on average only $20\%$ more time than a single model analysis. We then demonstrate that this technique can be used to accurately and efficiently quantify the support for one model over another, thereby assisting in Bayesian model selection.